How StackOverflow Scales with SQL Server

http://www.brentozar.com/archive/2011/11/how-stackoverflow-scales-sql-server-video/

by jswinghammer • 15 years ago

132 54 15 years ago

scottshea• 15 years ago

I love it when companies offer insight into their practices like this.

3am4• 15 years ago

I can't watch the video right now, but the idea that "Everybody's the DBA" is risky in general. I'm hugely in favor of developers writing and optimizing their own SQL, being able to create normalized schema (and know when it's worth the tradeoff to denormalize), how to read a query plan, and generally be as competent as a DBA. But... it's good to have one person who has the global view of the database for things like tuning extent sizes, selecting the optimal types of storage for various partions, doing reviews on the schema, capacity planning, etc, etc.

The right setup (IMO) of having a DBA in an operational role with developers that are highly proficient/self-sufficient is hard to get right and expensive enough that it probably isn't right for an early stage company. And a bad DBA can be a nightmare. So there are tradeoffs on both sides.

gaius6• 15 years ago

It's about moral authority. The only people touching production should be those whose pagers go off at 3am if it all goes hatstand. Everyone else is the peanut gallery.

3am4• 15 years ago

I love this explanation. Note my username :)

wahnfrieden• 15 years ago

Are you sure? Amazon is getting flak for that lately (in e.g. Yegge's rant), from having all of its developers on-call. You should be around to fix something you push if it breaks things, but that's not the same as being on-call at all hours for any given coworker's or manager's standard of "going hatstand".

jbert• 15 years ago

Agree, with the caveat that those who carry the 3am possibility are allowed to delegate at their own risk.

i.e. the real rule is "nothing happens to prod without the agreement of those at the sharp end".

That allows good devops fun to happen with fewer ops bottlenecks, if some devs can provide enough ops-chops to keep the real ops people happy.

[Also note in passing that such devs are likely to be in the "second wave" of 3am people if things go sufficiently pearish.]

Duff• 15 years ago

I think that with a modern database like SQL Server, the old-school "high priest" DBA is obsolete.

But... if you don't have someone dedicated to thinking about database issues, you need to treat database changes just like your code. It needs to be in a repository, it needs to be reviewed, and you need a change management regime.

From a anecdotal POV, I've noticed that many folks have a good process (or at least a consensus approach) to managing their code... but the database is often a red-headed stepchild that doesn't get the attention it deserves.

gaius6• 15 years ago

Same as with a modern language, the old-school "high priest" developer is obsolete.

reginaldo• 15 years ago

Thankfully, with Liquibase [1] and others like it, developers are starting to think about database change management.

I absolutely despise writing liquibase changesets, but I know how important they are.

[1] http://www.liquibase.org/

3am4• 15 years ago

Sure, I wasn't suggesting a high priest. Maybe more a 'faculty advisor' dba? Just someone keeping an eye on the higher level problems (having the right Nagios alerts set up, keeping an eye on how long reindexing is taking and scheduling downtime according, disaster recovery planning, putting in the POs for new drives). That's even if you treat writing queries and schema changes like part of the code, it's a dynamic system. But I completely agree on treating DB changes like code changes, I think DevOps is a good approach.

MartinCron2• 15 years ago

but the database is often a red-headed stepchild that doesn't get the attention it deserves

On my last three major projects, I have committed to devoting the proper level of attention to the database, with automatically building databases in some environments, scripted scheme changes and seed data loading as part of the mainline code base, etc. and I have found the difference to be immense in practical terms. I can move more quickly, more safely, and have a better quality of life as a developer.

One of the best returns on (effort) invested I have ever seen.

BrentOzar5• 15 years ago

You'll want to watch the video, because I explain that in more detail. Of course I'm not against change control or security, but if you want to scale a database, everybody involved has to have DBA-level skills. StackExchange's sysadmins and developers are all keenly aware of the database impact of what they do, and they aggressively try to minimize that impact. That's what I talk about on the video when I say "Everybody's the DBA." It's like saying, "Only you can prevent forest fires."

3am4• 15 years ago

What do you consider DBA level skills? Because you mention in one case people using SELECT * where they only needed a small subset of the fields. I'm all for your message that everyone should be aware of the database, but I don't consider that a DBA level mistake. I've just worked with some astonishingly good DBAs before, and I while I'm not a slouch it would be impossible to catch up with them without years of intensive study.

Anyway, I'm splitting hairs. Your presentation was great, and the people at SO are very impressive overall. Thanks for posting it.

andrewheins• 15 years ago

I love that these guys are so consistently open on their scaling strategies. Does anyone have any similar resources for Rails stacks?

moonboots2• 15 years ago

http://www.readwriteweb.com/cloud/2011/04/twitter-drops-ruby...

astrodust• 15 years ago

You should be so lucky as to have to take something to Twitter-scale.

I worry that if they went hand-rolled assembly, everyone would be jumping in that bandwagon.

donw• 15 years ago

I've downvoted you for providing a worthless answer.

The poster asked for similar resources on scaling Rails up; maybe something like https://github.com/blog/117-scaling-lesson-23742 or http://axonflux.com/building-and-scaling-a-startup would have been useful.

I sincerely doubt that the top-poster has Twitter-level problems or resources. He may have a Rails site that's crapping out under load, and needs resources for optimizing both application performance and database access.

"Just move everything to Java" is a meaningless answer. It doesn't help him at all, not unless he's got a year and change to retrain/rehire his engineers and completely rebuild his application.

krookoo• 15 years ago

> Does anyone have any similar resources for Rails stacks?

Of course; this:

Use Something Else.

krmmalik2• 15 years ago

I had no idea StackOverflow was using Microsoft SQL Server. I guess its just a learned response but i've always come to expect some large networking site to always be using a non-MS based solution.

I havent had a chance to watch the video, but i hope to later on. What is interesting, even with just the link is that for a website like StackOverflow that MS SQL is a viable solution.

We have been using SQL Server in our own company for our projects and i was really starting to get annoyed with it. I found it to be heavy on resources, slow to respond and lets not forget cost. I just completed a project that i have been working on for the last 4 months, and the majority of the work was within SQL server. One thing i learned was that its actually quite a powerful beast.

When used correctly and in the right way SQL is a very capable SQL solution. I'm glad that we decided to stick to SQL Server. There is a lot i learned about SQL Server in the last 4 months that i had no idea it was capable of.

silverbax885• 15 years ago

Big networks (banks, insurance companies) basically use one of two options most of the time. Oracle or SQL Server.

It only seems like it isn't when you read sites like Hacker News, where most of the posters are not working in big environments. Facebook is a large scale solution that doesn't use either, but their data management and caching is so bad nobody should be considering them as a best practice.

icebraining• 15 years ago

Google, Craigslist, Twitter, Wikipedia, Youtube and Netflix are just a few example which don't use Oracle or MS SQL Server either.

It's not about being big, banks and insurance companies simply have different requirements on how to access and store their data compared to most websites/services.

ja27• 15 years ago

Maybe it's because I worked so long for an IBM partner and dealt with IBM customers, but DB2 runs a lot of stuff at large Fortune 500 companies.

gaius6• 15 years ago

Oracle, Sybase, DB2, SQL Server and Informix. The last company I worked for had all of them.

  - Oracle and SQL Server: chosen by us
  - Sybase and Informix: chosen implicitly by choosing applications that preferred them[1]
  - DB2: Chosen by a company that we acquired

Oh and some MySQL and SQLite too.

[1] Yes they ran on Oracle too - but when 98% of a vendor's customers are on Informix or whatever, sometimes it's just less painful to go with the flow, so you have access to that community.

checoivan• 15 years ago

Some windows live properties backends use SqlServer as well and have thousands of SQL instances running, so it does scale.

http://windowsteamblog.com/windows_live/b/windowslive/archiv...

timwiseman• 15 years ago

I found it to be heavy on resources, slow to respond and lets not forget cost

What are you comparing it to?

Admittedly I am biased since my day job is a SQL Server DBA, but I have tried several other options and think SQL Server tends to stack up quite well.

I have found it generally more user friendly, easier to work with, and cheaper than Oracle (though Oracle does seem to have an advantage in certain types of partitioning). I rather like MySql for certain types of projects, but generally find SQL Server easier to maintain for large projects.

I have only dabbled with NoSQL options, but my general opinion is that for certain problem sets, they are great. However, when ACID is even remotely desireable they are not an option and for certain other tasks they are less desirable.

So, I think that which type of database you use depends largely on the project, but SQL Server tends to stand up quite well for a wide array of projects.

krmmalik2• 15 years ago

I was comparing to MySQL and PostgresSQL, although i can understand that it may not necessarily be a like for like comparison.

Ultimately, i agree with what you are saying and my opinion of SQL server has improved in the last few months. I really like its flexibility.

mkramlich• 15 years ago

one or both of the main people behind SO are "Microsoft stack" guys, so I'm not surprised.

blhack• 15 years ago

"Don't worry the cloud takes care of all of this!"

This is horrifying. Are there really people that think this way?

BrentOzar5• 15 years ago

Sadly, yes - obviously it's not this site's target audience, but a lot of PHBs out there are buying into it hook, line, and sinker.

wmwong• 15 years ago

This wasn't so much about fine tuning SQL Server itself, but breaking traditional thinking. The 5 rules he presents originally are all replaced in the end.

  - Everybody's the DBA
  - Do what it takes to get what you want
  - Tune later, cache & separate now
  - NewEgg your way out of problems
  - Share for great good

For each point, he argues why the old rule no longer applies and what the new solution is.

I felt a lot of the presentation was about tuning SQL Server without tuning SQL Server: caching, leave full-text searching to Apache Lucene (because it's not querying), and using SSDs to speed up performance without having to touch any code.

newegg• 15 years ago

Interesting that buying hardware written as "NewEgg your way out of a problem" as Newegg uses IIS and quite likely SQL Server.

BrentOzar5• 15 years ago

Yep, absolutely, you nailed it. There's a gazillion presentations out there about tuning databases, but that only takes you so far. I wanted to show that you need to take a step back before you go into query tuning details.

arthurprs• 15 years ago

Very good stuff.

revetkn• 15 years ago

The other part of this is throwing out the ORM in places where performance matters and replacing it with a simple utility, Dapper, that maps SQL to objects: http://samsaffron.com/archive/2011/03/30/How+I+learned+to+st...

mwsherman• 15 years ago

It comes down to two things: economics and computer science, in that order.

For most companies, diving deep into your data persistence layer is probably not worth it in the beginning. Your devs need to work on features. (MS SQL does pretty well untuned.) This is the economics part.

The computer science part comes in when you are doing enough traffic that 200 hours of dev time toward a 10% performance improvement becomes good economics. Then you dig into the data layer (and every other layer) and start counting milliseconds.

Which is what we did at Stack O by using Dapper and renting Brent Ozar. :)

BrentOzar5• 15 years ago

RENTING, hahaha, I never thought of myself that way. I like it. I'm going to use a "For Rent" sign around my neck in an upcoming blog post!

How StackOverflow Scales with SQL Server

Related Stories